Memory · Skills · Agents: Building an AI Data Analysis Team That Grows Over Time

Imagine this scenario:

On Monday morning, you ask an AI assistant: "How were sales in East China last week?" It checks the data and gives you the result. You add: "From now on, for this kind of sales data, please include month-over-month comparison. Our team prefers to see that." The AI says okay.

On Wednesday, you ask: "How are things in South China this week?"

Then you discover—it gives you another plain number, with no comparison, no context. As if that Monday conversation never happened.

This isn't a story about "poor memory"—it's the true portrayal of most current AI data analysis tools.

Each conversation starts from scratch, like an employee who's always in their first week: you repeatedly explain business terminology, correct output formats, and tell it "we don't look at absolute values, we look at trends." It's indeed smart, but this intelligence can't accumulate.

AskTable is changing this.

Through the synergy of three core capabilities—Memory System, Skill System, and Industry Agents—AskTable is transforming AI from a "always a newcomer" tool into a continuously growing data team partner.

---

I. The Root Problem: Why General AI Can't Do Data Analysis Well

1.1 The "Newcomer" Dilemma

In data analysis scenarios, a general AI assistant faces three fundamental shortcomings:

1.2 Solution: Three-in-One

AskTable's approach isn't about patching individual points, but building three mutually supportive capability pillars at the system level:

These three capabilities aren't simply叠加—they form a positive feedback loop: memory makes agents understand users better, Skills make agents more capable, and agent usage in turn generates new memory and skill optimization.

Let's break down these three pillars one by one.

---

II. Memory: Letting AI Stop Forgetting

2.1 What the Memory System Solves

The core problem the memory system solves is: keeping AI coherent across different sessions.

This isn't simply "saving chat history," but selectively extracting, storing, and retrieving information valuable to users:

• •User preferences ("I prefer to see MoM data")
• •Business context ("Our Q1 target is 50 million")
• •Corrections and feedback ("This metric shouldn't be broken down by day, view it monthly")
• •Historical patterns ("Users usually look at last month's summary at the beginning of the month")

2.2 Technical Architecture

The memory system is built on mem0 + Qdrant, using a read-write separation design:

Key Design Decisions:

1. •

   Read-Write Separation: Search executes synchronously at the start of a request, Add executes asynchronously after the reply completes, ensuring no impact on response speed.

2. •

   Agent-Level Isolation: The memory isolation granularity is at the Data Agent level, not user level. This means all users under the same Data Agent (such as "Sales Analysis Assistant") share memory, forming team-level collective memory.

3. •

   Protocol Abstraction Layer: The memory system is defined through Protocol interfaces, with mem0 as the current implementation. This ensures seamless switching to other memory solutions in the future without affecting upper-layer logic.

4. •

   LLM-Driven Memory Extraction: Instead of simply saving conversation text, structured memory information (preferences, facts, corrections) is extracted from conversations through LLM, deduplicated, and stored in a vector database.

2.3 Memory Lifecycle

Memory isn't simply "stored" and "retrieved"—it has complete lifecycle management:

Three Types of Memory Extraction:

This classification ensures the memory system doesn't grow infinitely, while important long-term preferences are retained.

2.4 Protocol Abstraction Layer Design

The memory system is defined through Protocol interfaces, not directly dependent on mem0's concrete implementation:

class MemoryProvider(Protocol):
    """Memory provider protocol—ensuring replaceability"""

    async def search(
        self,
        query: str,
        agent_id: str,
        limit: int = 5,
    ) -> list[MemoryEntry]:
        """Retrieve relevant memories based on query"""
        ...

    async def add(
        self,
        messages: list[Message],
        agent_id: str,
    ) -> list[MemoryEntry]:
        """Extract and store new memories from conversations"""
        ...

    async def get_all(self, agent_id: str) -> list[MemoryEntry]:
        """Get all memories under an Agent"""
        ...

    async def delete(self, memory_id: str) -> bool:
        """Delete specified memory"""
        ...


# Current implementation: mem0
class Mem0MemoryProvider:
    def __init__(self, config: Mem0Config):
        self.memory = AsyncMemory.from_config(config)

    async def search(self, query, agent_id, limit=5):
        results = await self.memory.search(
            query=query,
            user_id=agent_id,
            limit=limit,
        )
        return [self._to_entry(r) for r in results]

    # ... other method implementations

This means if there's a need to switch to a self-developed memory solution or other open-source solution in the future, only the same Protocol interface needs to be implemented, with no modifications to upper-layer code.

2.5 Actual Effect

Here's a comparison:

Without the memory system:

Conversation A (Monday):
  User: Sales in East China
  AI: 12 million
  User: From now on, include YoY and MoM for this kind of data
  AI: Okay

Conversation B (Wednesday):
  User: Sales in South China
  AI: 8 million                          ← No YoY or MoM

With the memory system:

Conversation A (Monday):
  User: Sales in East China
  AI: 12 million
  User: From now on, include YoY and MoM for this kind of data
  AI: Okay                    → Memory system extracts and stores:
                                 "User preference: Sales data needs YoY and MoM"

Conversation B (Wednesday):
  User: Sales in South China
  AI: 8 million, YoY +12%, MoM -3%     ← Automatically applies preference
      Note: MoM decline mainly affected by XX category

The meaning of the memory system: AI no longer needs you to teach it repeatedly. It沉淀 every valuable piece of information from each interaction and automatically calls it up in the next conversation.

For technical details on the memory system, recommended reading: AskTable Permanent Memory System: Letting AI Remember Every Conversation

---

III. Skill System: On-Demand Professional Capabilities

3.1 From "Knows Everything" to "Knows How to Call Professional Knowledge"

If the memory system solves AI's "experience accumulation" problem, then the Skill system solves the "professional capability" problem.

Traditional Agent approaches stuff all capabilities into system prompts or tool lists. This creates two problems:

1. •Context pollution: Each request carries many tool definitions, increasing reasoning costs and error probability
2. •Rigid capabilities: Adding or modifying capabilities requires code changes and redeployment

AskTable's Skill system uses a more elegant approach: project-level skill library + hot-swappable loading.

3.2 Skill System Architecture

Core Design Points:

• •Hot-swappable: Skills are plain text instructions, stored in a database, no code changes needed to add or modify
• •Dynamic loading: Through the activate_skill() tool, Agents can activate skills on-demand during conversations, expanding their capability boundaries
• •Three-layer priority: explicit > Agent > Project, ensuring flexibility and default capability balance

3.3 Three-Layer Loading Mechanism Explained

The three-layer loading mechanism is one of the core innovations in the Skill system. It ensures that in different scenarios, Agents can get the most appropriate skill set:

The meaning of this design: neither forcing Agents to carry all skills every time (avoiding context pollution), nor sacrificing capability availability when needed.

3.4 Overview of 11 Built-in Skills

AskTable has 11 built-in professional data analysis skills, covering the complete analysis chain from data exploration to report generation:

These 11 skills are like a analysis team's "standard arsenal"—each skill is a core capability of a professional data analyst.

For in-depth analysis of the Skill system, recommended reading: AskTable Skill System: Letting AI Agents Call Professional Capabilities On Demand

---

IV. Agents: Industry Experts Combining Skills

4.1 From Skills to Agents: 1 + 1 > 2

With memory and skills, the next step is how to organize them into a "knowledgeable" expert. This is the value of industry agents.

Agents aren't simply a stack of skills, but an encapsulation of skill combinations + industry knowledge + analysis frameworks for specific business scenarios.

4.2 9 Major Industry Agents

AskTable comes with 9 pre-built industry agents, covering the most common data analysis scenarios:

4.3 How Agents Work

The core value of agents is understanding your business context and automatically selecting the right skill combination.

Take an e-commerce operations scenario as an example:

The reason agents can do this is because they:

• •Understand key metrics and analysis frameworks for this industry
• •Can activate relevant analysis skills on demand
• •Remember user preferences and historical context

For complete introduction to built-in skills and agents, recommended reading: AskTable Built-in Skills and Agents: Out-of-the-Box Data Analysis Capabilities

---

V. Three-in-One: Complete Workflow of Collaborative Work

5.1 Synergistic Relationship Between the Three

Memory, Skills, and Agents don't work independently—they form a mutually reinforcing positive feedback loop:

5.2 Complete Workflow

Let's look at a complete usage scenario to see how the three work together:

Scenario: A retail enterprise operations director uses AskTable for monthly operations analysis

═══════════════════════════════════════════════════════════════

Step 1: User Initiates Request
───────────────────────────────────────────────────────────────
User: "Show me last month's operations situation"

═══════════════════════════════════════════════════════════════

Step 2: Memory System Involvement (Read)
───────────────────────────────────────────────────────────────
Memory search discovers:
  ✓ User's Data Agent is "Executive Data Assistant"
  ✓ User preference: focuses on revenue, profit margin, YoY and MoM
  ✓ Last correction: "Use previous month for MoM, not moving average"
  ✓ Business context: Q1 revenue target 50 million, currently at 62%

→ This information is injected into the system prompt

═══════════════════════════════════════════════════════════════

Step 3: Agent Selects Skills
───────────────────────────────────────────────────────────────
The "Executive Data Assistant" agent, after analyzing user needs, decides to activate:

  activate_skill("Cycle Analysis")       → Identify last month's overall trend
  activate_skill("Comparative Analysis")       → YoY, MoM, target achievement
  activate_skill("Anomaly Detection")       → Check for any anomalous fluctuations
  activate_skill("Drill-Down Metrics")       → Locate main growth/decline sources
  activate_skill("Business Language Generation")   → Generate management-readable conclusions

═══════════════════════════════════════════════════════════════

Step 4: Execute Analysis
───────────────────────────────────────────────────────────────
AI executes skills in sequence to generate structured analysis:

  📊 Last Month Operations Summary
  ━━━━━━━━━━━━━━━━━━━━━━━━
  Revenue: 18.5M | YoY +15% | MoM +8%
  Profit Margin: 18.5% | YoY +1.2pp | MoM -0.3pp
  Q1 Target Achievement: 62% (on track)

  📈 Key Findings:
  01. Revenue growth mainly driven by new products in East China (contribution rate 65%)
  02. Profit margin MoM slight decline, mainly due to raw material cost increase
  03. South China performance abnormally low (2.1 standard deviations below mean)

  ⚠️ Risk Alert:
  South China declining for two consecutive weeks, needs attention

═══════════════════════════════════════════════════════════════

Step 5: User Interaction and Feedback
───────────────────────────────────────────────────────────────
User: "That South China anomaly, help me focus on which stores have the problem"

→ Memory system records user's focus
→ Agent activates "Drill-Down Metrics" and "Anomaly Detection" skills
→ Generate detailed analysis

═══════════════════════════════════════════════════════════════

Step 6: Memory System Write (Async)
───────────────────────────────────────────────────────────────
After conversation ends, memory system async executes:
  ✓ Extract: User focuses on anomalous stores
  ✓ Extract: User continuously tracking South China performance
  ✓ Update: South China as current focus area

→ Next conversation, AI will proactively focus on South China

═══════════════════════════════════════════════════════════════

5.3 The Essence of Growth: Positive Feedback Loop

The core advantage of this system is that it's a self-reinforcing positive feedback loop:

Every conversation accumulates value for the next. This is the true meaning of a "team that grows."

---

VI. Real Case: From "Tool" to "Colleague" Transformation

6.1 Case Background

A mid-sized retail enterprise (annual revenue ~200M), changes after three months of using AskTable:

6.2 Usage Journey

Week 1: Tool Stage

Operations Director Li Ming used AskTable for the first time, treating it as "a chatbot that can query data."

Li Ming: "Sales by region last month"
AI:      "East 5.8M, South 3.2M, North 4.1M, Southwest 1.9M"
Li Ming: (thinking) Okay, data is correct, but too simple

At this point, AI is similar to an intern who can write SQL.

Week 2 to Month 1: Learning Stage

Li Ming started giving AI feedback:

Li Ming: "From now on, include YoY and MoM for this kind of data"
AI:      "Okay"

Li Ming: "Why is South China so bad this month?"
AI:      (calls Anomaly Detection + Drill-Down analysis)
       "South China 3.2M, MoM down 22%, YoY down 8%
        Main drag: Guangzhou Tianhe store (-45%), Shenzhen Nanshan store (-31%)
        Common factor: both stores had renovation downtime"
Li Ming: (thinking) This analysis is pretty good

Memory system started working—Li Ming's preferences, focus areas, and analysis habits were recorded.

Month 2 to 3: Colleague Stage

Li Ming: "This month's operations situation"
AI:      "📊 March Operations Express
         ━━━━━━━━━━━━━━━
         Revenue 18.5M, YoY +15%, MoM +8%
         Profit Margin 18.5%, YoY +1.2pp
         Q1 Target Achievement 62%, on track

         📈 Highlights:
         East China new product series drove 12% revenue growth

         ⚠️ Attention:
         South China still recovering, but MoM up 5%
         (Tianhe store you focused on last month is back to normal operations)

         Should I generate a detailed monthly operations report?"

Li Ming: (thinking) This is already half an operations assistant

At this point, AI:

• •Remembers Li Ming cares about which metrics
• •Knows South China is an area he continuously tracks
• •Proactively calls multiple analysis skills to generate structured reports
• •Can even anticipate his next needs

This is the transformation from "tool" to "colleague."

6.3 Quantified Changes

---

VII. Technical Details: How the Three Truly Collaborate

7.1 How Memory Affects Skill Selection

Information stored in the memory system directly affects agent skill selection:

# Pseudocode: Memory-enhanced skill selection
async def select_skills_with_memory(agent, user_query):
    # 1. Retrieve relevant context from memory system
    memories = await memory.search(user_query, agent_id=agent.id)

    # 2. Inject memory into prompt
    enhanced_prompt = build_prompt(
        base_prompt=agent.system_prompt,
        memories=memories,  # Contains user preferences and historical focus
    )

    # 3. Agent selects skills based on enhanced prompt
    selected_skills = await agent.decide_skills(
        query=user_query,
        prompt=enhanced_prompt,
    )

    # 4. Activate selected skills
    for skill in selected_skills:
        await agent.activate_skill(skill)

    return selected_skills

7.2 How Skills Generate New Memory

Skill execution results and user feedback become new memories:

User asks: "Show me last month's operations"
  │
  ├── AI calls multiple skills to generate report
  │
  ├── User feedback: "I want to see South China data weekly"
  │     │
  │     └── Memory system extracts:
  │         "User wants to track South China data weekly"
  │
  └── Next time user asks about operations
        │
        └── AI proactively includes South China weekly analysis

7.3 How Agents Coordinate the Three

The agent is the "brain" of the entire system, responsible for:

┌─────────────────────────────────────────────────┐
│              Agent Scheduling Flow                       │
│                                                 │
│  Receive user request                                    │
│       │                                         │
│       ▼                                         │
│  ┌─────────────────┐                            │
│  │  Understand user intent      │                            │
│  └────────┬────────┘                            │
│           │                                     │
│           ▼                                     │
│  ┌─────────────────┐                            │
│  │  Retrieve memory context    │  ← Read memory                  │
│  └────────┬────────┘                            │
│           │                                     │
│           ▼                                     │
│  ┌─────────────────┐                            │
│  │  Select related skills      │  ← Use Skill                    │
│  └────────┬────────┘                            │
│           │                                     │
│           ▼                                     │
│  ┌─────────────────┐                            │
│  │  Execute analysis and reply   │                            │
│  └────────┬────────┘                            │
│           │                                     │
│           ▼                                     │
│  ┌─────────────────┐                            │
│  │  Async write new memory       │  ← Write memory                  │
│  └────────┬────────┘                            │
│                                                 │
└─────────────────────────────────────────────────┘

---

VIII. Why This Architecture Makes Sense

8.1 Rethinking the Data Analysis Paradigm

Traditional data analysis tools are essentially "query engines"—you input questions, it returns data. Each interaction is independent, with no accumulation, no growth.

AskTable's approach is to transform AI data analysis into a **"growing team":

8.2 Three Core Differences

8.3 Future Scalability

The design philosophy of this architecture is openness and extensibility:

1. •Memory system's Protocol abstraction: mem0 is just one implementation, can be replaced anytime
2. •Skill hot-swapping: Adding skills doesn't require code changes, supports AI-assisted generation
3. •Agent customizability: Enterprises can create exclusive industry agents based on their needs

This means AskTable's capability boundary isn't fixed, but continuously expands with use.

---

IX. Summary

"Good AI tools don't get smarter—they get to know you better."

Reviewing AskTable's three core capabilities:

1. •

   Memory System: Letting AI stop forgetting. Through mem0 + Qdrant's read-write separation design, accumulating team-level collective memory at the Data Agent level.

2. •

   Skill System: Letting AI have professional capabilities. Project-level skill library + hot-swappable loading + 11 built-in skills, covering the complete analysis chain from anomaly detection to report generation.

3. •

   Industry Agents: Letting AI understand the business. 9 major agents combine skills for specific business scenario analysis capabilities, out of the box.

The three work together, forming a continuously evolving positive feedback loop—memory makes agents understand users better, skills make agents more capable, and every agent use generates new memory and skill optimization.

This isn't about building a better query tool—it's about building a data analysis team that grows.

---

Extended Reading

• •AskTable Permanent Memory System: Letting AI Remember Every Conversation — Memory system technical details
• •AskTable Skill System: Letting AI Agents Call Professional Capabilities On Demand — Skill system architecture design
• •AskTable Built-in Skills and Agents: Out-of-the-Box Data Analysis Capabilities — Complete introduction to 11 skills and 9 agents
• •AskTable AI Agent Workflow: From Architecture to Implementation — In-depth Agent infrastructure analysis